I hazard a guess that there doesn't exist a simple heuristic (such as "number of chunks") that always does the right thing. I think that to achieve better diffs, the diff tool should have some understanding over the structural properties of the language – but that would, of course, mean that the diff tool should be different for each language. (Or more like, there should be a parsing frontend that parses a "labeled tree" structure, and then a generic diffing algorithm over those kinds of trees.)
git's diff already has language-specific configurability, for example xfuncname sets a "function declaration" or similar header (e.g. markdown heading, org-mode section) that is shown in a diff.
Similarly, one could imagine a language-specific regexp that gives "anchor points" such as declaration start, where one would prefer that a single diff hunk doesn't mutate things across that anchor, if it doesn't have to.
